TeloaAI-Native Team Studio

Back to the marketplace · Skill

Code change cleanup

Review recent code changes for reuse, quality, efficiency and abstraction; combine findings and verify fixes. Derived from the Hermes Agent skill: efficiency findings must state the work per call before and after the fix, suggested fixes carry a behavior-equivalence check, any angle may report "no material findings", and an output-budget note is added.

EngineeringDev Tools

Copy this sentence and send it to any colleague in Teloa:
In Teloa, open Marketplace, search for "Code change cleanup" and add it (Skill hermes.simplify-code).

Source
Teloa official · Derived from Hermes Agent (NousResearch/hermes-agent)

Change list (8)

Adapted (2)

  • Adds a prominent change notice below the title, "Derived work: modified by Teloa from NousResearch/hermes-agent@a7c080ca…", pointing to MODIFICATIONS.md.

    Why: MIT does not require a change notice, but Teloa requires derived files to state their changes; visible text is used instead of an HTML comment, which security scans would flag as hidden content.

    SKILL.md · below the title HSC-M08 · View original

  • The frontmatter keeps only `name` and `description`; `version` (earlier changed from `1.1.0` to `1.1.0-teloa.1` so it would not pass for the original 1.1.0; the entry version now applies), `author`, `license`, `platforms` and `metadata.hermes` (tags, related_skills) are removed. The body is unchanged.

    Why: Skills stored in the Teloa marketplace may only have the single-line `name` and `description` keys in their frontmatter; authorship and license information stay in this entry's original source, the root LICENSE and MODIFICATIONS.md.

    SKILL.md · frontmatter HSC-M09 · View original

Added (5)

  • The finding format adds a "behavior check (why the fix keeps outputs, iteration order, side effects and error behavior identical)"; CAREFUL/RISKY fixes without it are marked `confidence: low`.

    Why: The original relies on running tests after applying fixes in Phase 3; in report-only runs nothing required an equivalence argument.

    SKILL.md · §Phase 2, finding format (L110) HSC-M02 · View original

  • Adds two rules right next to the finding format that apply to every finding: an efficiency finding must state the work per call before and after the fix, and if they are equal it is not an efficiency finding; any angle may return `no material findings`, an empty result is valid and a made-up one is not.

    Why: These rules did not exist; next to the output format they are in view while each finding is written.

    SKILL.md · §Phase 2, after the finding format (after L113) HSC-M03 · View original

  • Adds: an angle may have no material findings and reports `no material findings`, without relabelling another angle's findings or inventing gains to fill the slot; duplicated code costs maintenance, not CPU, unless the same input is really processed several times on the same execution path; without call sites that prove repeated work, efficiency defaults to `no material findings (call sites not provided)`.

    Why: The original had no way to return an empty result, and callers often require a result for all four angles, which together create pressure to fill the slot; this gives a ready, valid answer.

    SKILL.md · §Phase 2, end of "Tell each reviewer to" list (after L121-122) HSC-M04 · View original

  • Adds a self-check pass: after the sequential fallback covers all four angles, re-check only the efficiency findings (does the work per call really differ, and which line proves it) and drop those that do not hold.

    Why: Without delegation one model covers all four angles in turn, so the error-prone efficiency conclusions get a separate check.

    SKILL.md · §Phase 2, "No delegation available?" (L91-97) HSC-M05 · View original

  • Adds "Reasoning budget": models that think before answering may need more than 8k output tokens for an inline four-angle review; a truncated answer is not a review, so raise the output limit or narrow the diff and rerun, and never act on a partial result.

    Why: In testing, a 4096-token output limit let the model exhaust the budget while thinking and produce no answer.

    SKILL.md · §Pitfalls (after L245) HSC-M06 · View original

Improved (1)

  • Changes "For each, give the concrete fix and why it's faster or safer." to: give the concrete fix and a measurable basis, citing the call sites (file:line) that actually repeat work within one request or run and stating the work per call before and after the fix (for example: before, one DNS lookup per outgoing message; after, one per batch); if before and after are the same it is not an efficiency finding and is dropped or handed to the Quality angle; sharing code between separately called functions does not by itself reduce runtime work (two CLI subcommands sharing an argument parser still parse once each), only repeated work on the same execution path counts; for silent failures and TOCTOU, say which error or race is currently hidden and what would expose it, as the basis for "safer".

    Why: The original asks every efficiency finding to explain why it is faster, which invites made-up gains when the input has no efficiency problem. The new wording asks for a checkable basis and keeps the "or safer" argument. Comparative testing found no detectable difference between the two versions in made-up findings; this change rests on the correctness of the instruction.

    SKILL.md · §Phase 2, Reviewer 3 (Efficiency), last sentence (L166-167) HSC-M01 · View original

The other file matches the original (verified)

GitHub
Resource file: catalog/skills/hermes.simplify-code.json
Files kept by the marketplace: artifacts/skills/hermes.simplify-code/1.1.0/
Original source (locked version): GitHub NousResearch/hermes-agent@a7c080c
License
MIT
License files: LICENSE
Review
Reviewed by Teloa on 2026-09-28
Compatibility
Needs configuration · Teloa >=0.2.0-alpha.7 · DSH 0.1.7-rc.1
  • Use on an explicitly provided recent change (diff or file scope). Needs Git and the bash/read/edit tools; without Hermes delegation it falls back to a single inline pass.
  • Callers must allow any angle to return `no material findings` and must not require a finding for every angle.
  • Check efficiency conclusions by hand: in testing, qwen3:8b described a maintenance benefit as a runtime gain in about one in five samples. qwen3:4b and 8b were tested; larger models were not.
  • For models that think before answering, set the output limit to at least 8k tokens; a truncated answer is not a review. It does not install dependencies, commit or push on its own.
Permissions and requirements
Tools: bash, read, edit
Required software: git
Uses the network: no
Data flow
The files come with Teloa; nothing is downloaded when you add it. You can browse without signing in. Page views are counted with Cloudflare Web Analytics, which uses no cookies and does not record who you are. Data used for sign-in and reviews: privacy and community rules

Skill hermes.simplify-code · version 1.1.0

Report a problem